ECG-NAT: A Self-supervised Neighborhood Attention Transformer for Multi-lead Electrocardiogram Classification
Researchers have developed ECG-NAT, a self-supervised Neighborhood Attention Transformer designed for multi-lead electrocardiogram classification. This model uses a two-stage approach, beginning with generative pretraining on unlabeled ECG data to learn robust representations, followed by discriminative fine-tuning with a dual-loss function. ECG-NAT's hierarchical attention mechanism efficiently captures both fine-grained beat morphology and broader rhythm patterns, achieving 88.1% accuracy with only 1% labeled data, making it effective in low-resource scenarios. AI
IMPACT Introduces a novel self-supervised learning approach for ECG classification, improving accuracy in low-data scenarios.